METHOD OF INTELLIGENT DETECTION OF PROPAGANDA TECHNIQUES BY SIGNS USING MACHINE LEARNING

The article is devoted to creation and approbation of a method of intelligent detection of propaganda techniques by signs using machine learning, which allows to transform input data in the form of analysis text and controlled machine learning models into output data containing numerical evaluations of manifestation of each propaganda technique and marked text with visual analytics the presence of detected propaganda signs. A study was conducted, which allows to identify 17 main propaganda techniques: "Appeal to fear-prejudice", "Causal Oversimplification", "Doubt", "Exaggeration", "Flag-Waving", "Labeling", "Loaded Language", "Minimisation ", "Name Calling", "Repetition", "Appeal to Authority", "Black and White Fallacy", "Reductio ad hitlerum", "Red Herring", "Slogans", "Thought terminating Cliches", "Whataboutism". Study compared three approaches that are most often used for similar problems: traditional machine learning approach, approach based on recurrent neural networks, and approach based on transformer models. Traditional machine learning approaches performed worse, reaching accuracies between 0.60 and 0.67, because they are unable to take into account context, which is important for detecting propaganda techniques. Recurrent neural networks showed better results than traditional approaches, achieving accuracy from 0.66 to 0.80, but they still had difficulty handling long dependencies. The highest results were achieved by an approach based on the transformer neural network model, which uses self-awareness mechanisms that allow each element of sequence to directly interact with all other elements. This ensures the effective capture of long-term dependencies, which is characteristic of propaganda techniques. The results obtained on the basis of approach based on neural network transformer model ensured the detection of various propaganda techniques with minimum accuracy of 79.03% and maximum accuracy of 97.07%. At the same time, minimum accuracy values were obtained for "Whataboutism" technique, and maximum accuracy values were obtained for "Loaded Language" technique. The obtained accuracy values are better than all known analogues in detecting propaganda, regardless of methods used, which opens up opportunities for effective practical use of proposed method. Keywords: propaganda detecting, NLP, RNN, BERT, propaganda techniques, propaganda signs.